mm/z3fold.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * z3fold.c
   4  *
   5  * Author: Vitaly Wool <vitaly.wool@konsulko.com>
   6  * Copyright (C) 2016, Sony Mobile Communications Inc.
   7  *
   8  * This implementation is based on zbud written by Seth Jennings.
   9  *
  10  * z3fold is an special purpose allocator for storing compressed pages. It
  11  * can store up to three compressed pages per page which improves the
  12  * compression ratio of zbud while retaining its main concepts (e. g. always
  13  * storing an integral number of objects per page) and simplicity.
  14  * It still has simple and deterministic reclaim properties that make it
  15  * preferable to a higher density approach (with no requirement on integral
  16  * number of object per page) when reclaim is used.
  17  *
  18  * As in zbud, pages are divided into "chunks".  The size of the chunks is
  19  * fixed at compile time and is determined by NCHUNKS_ORDER below.
  20  *
  21  * z3fold doesn't export any API and is meant to be used via zpool API.
  22  */
  23
  24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  25
  26 #include <linux/atomic.h>
  27 #include <linux/sched.h>
  28 #include <linux/cpumask.h>
  29 #include <linux/list.h>
  30 #include <linux/mm.h>
  31 #include <linux/module.h>
  32 #include <linux/page-flags.h>
  33 #include <linux/migrate.h>
  34 #include <linux/node.h>
  35 #include <linux/compaction.h>
  36 #include <linux/percpu.h>
  37 #include <linux/mount.h>
  38 #include <linux/pseudo_fs.h>
  39 #include <linux/fs.h>
  40 #include <linux/preempt.h>
  41 #include <linux/workqueue.h>
  42 #include <linux/slab.h>
  43 #include <linux/spinlock.h>
  44 #include <linux/zpool.h>
  45 #include <linux/magic.h>
  46 #include <linux/kmemleak.h>
  47
  48 /*
  49  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  50  * adjusting internal fragmentation.  It also determines the number of
  51  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  52  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  53  * in the beginning of an allocated page are occupied by z3fold header, so
  54  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  55  * which shows the max number of free chunks in z3fold page, also there will
  56  * be 63, or 62, respectively, freelists per pool.
  57  */
  58 #define NCHUNKS_ORDER   6
  59
  60 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  61 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  62 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  63 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  64 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  65 #define NCHUNKS         (TOTAL_CHUNKS - ZHDR_CHUNKS)
  66
  67 #define BUDDY_MASK      (0x3)
  68 #define BUDDY_SHIFT     2
  69 #define SLOTS_ALIGN     (0x40)
  70
  71 /*****************
  72  * Structures
  73 *****************/
  74 struct z3fold_pool;
  75 struct z3fold_ops {
  76         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  77 };
  78
  79 enum buddy {
  80         HEADLESS = 0,
  81         FIRST,
  82         MIDDLE,
  83         LAST,
  84         BUDDIES_MAX = LAST
  85 };
  86
  87 struct z3fold_buddy_slots {
  88         /*
  89          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
  90          * be enough slots to hold all possible variants
  91          */
  92         unsigned long slot[BUDDY_MASK + 1];
  93         unsigned long pool; /* back link */
  94         rwlock_t lock;
  95 };
  96 #define HANDLE_FLAG_MASK        (0x03)
  97
  98 /*
  99  * struct z3fold_header - z3fold page metadata occupying first chunks of each
 100  *                      z3fold page, except for HEADLESS pages
 101  * @buddy:              links the z3fold page into the relevant list in the
 102  *                      pool
 103  * @page_lock:          per-page lock
 104  * @refcount:           reference count for the z3fold page
 105  * @work:               work_struct for page layout optimization
 106  * @slots:              pointer to the structure holding buddy slots
 107  * @pool:               pointer to the containing pool
 108  * @cpu:                CPU which this page "belongs" to
 109  * @first_chunks:       the size of the first buddy in chunks, 0 if free
 110  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
 111  * @last_chunks:        the size of the last buddy in chunks, 0 if free
 112  * @first_num:          the starting number (for the first handle)
 113  * @mapped_count:       the number of objects currently mapped
 114  */
 115 struct z3fold_header {
 116         struct list_head buddy;
 117         spinlock_t page_lock;
 118         struct kref refcount;
 119         struct work_struct work;
 120         struct z3fold_buddy_slots *slots;
 121         struct z3fold_pool *pool;
 122         short cpu;
 123         unsigned short first_chunks;
 124         unsigned short middle_chunks;
 125         unsigned short last_chunks;
 126         unsigned short start_middle;
 127         unsigned short first_num:2;
 128         unsigned short mapped_count:2;
 129         unsigned short foreign_handles:2;
 130 };
 131
 132 /**
 133  * struct z3fold_pool - stores metadata for each z3fold pool
 134  * @name:       pool name
 135  * @lock:       protects pool unbuddied/lru lists
 136  * @stale_lock: protects pool stale page list
 137  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 138  *              buddies; the list each z3fold page is added to depends on
 139  *              the size of its free region.
 140  * @lru:        list tracking the z3fold pages in LRU order by most recently
 141  *              added buddy.
 142  * @stale:      list of pages marked for freeing
 143  * @pages_nr:   number of z3fold pages in the pool.
 144  * @c_handle:   cache for z3fold_buddy_slots allocation
 145  * @ops:        pointer to a structure of user defined operations specified at
 146  *              pool creation time.
 147  * @zpool:      zpool driver
 148  * @zpool_ops:  zpool operations structure with an evict callback
 149  * @compact_wq: workqueue for page layout background optimization
 150  * @release_wq: workqueue for safe page release
 151  * @work:       work_struct for safe page release
 152  * @inode:      inode for z3fold pseudo filesystem
 153  *
 154  * This structure is allocated at pool creation time and maintains metadata
 155  * pertaining to a particular z3fold pool.
 156  */
 157 struct z3fold_pool {
 158         const char *name;
 159         spinlock_t lock;
 160         spinlock_t stale_lock;
 161         struct list_head *unbuddied;
 162         struct list_head lru;
 163         struct list_head stale;
 164         atomic64_t pages_nr;
 165         struct kmem_cache *c_handle;
 166         const struct z3fold_ops *ops;
 167         struct zpool *zpool;
 168         const struct zpool_ops *zpool_ops;
 169         struct workqueue_struct *compact_wq;
 170         struct workqueue_struct *release_wq;
 171         struct work_struct work;
 172         struct inode *inode;
 173 };
 174
 175 /*
 176  * Internal z3fold page flags
 177  */
 178 enum z3fold_page_flags {
 179         PAGE_HEADLESS = 0,
 180         MIDDLE_CHUNK_MAPPED,
 181         NEEDS_COMPACTING,
 182         PAGE_STALE,
 183         PAGE_CLAIMED, /* by either reclaim or free */
 184 };
 185
 186 /*
 187  * handle flags, go under HANDLE_FLAG_MASK
 188  */
 189 enum z3fold_handle_flags {
 190         HANDLES_NOFREE = 0,
 191 };
 192
 193 /*
 194  * Forward declarations
 195  */
 196 static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
 197 static void compact_page_work(struct work_struct *w);
 198
 199 /*****************
 200  * Helpers
 201 *****************/
 202
 203 /* Converts an allocation size in bytes to size in z3fold chunks */
 204 static int size_to_chunks(size_t size)
 205 {
 206         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 207 }
 208
 209 #define for_each_unbuddied_list(_iter, _begin) \
 210         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 211
 212 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
 213                                                         gfp_t gfp)
 214 {
 215         struct z3fold_buddy_slots *slots;
 216
 217         slots = kmem_cache_zalloc(pool->c_handle,
 218                                  (gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE)));
 219
 220         if (slots) {
 221                 /* It will be freed separately in free_handle(). */
 222                 kmemleak_not_leak(slots);
 223                 slots->pool = (unsigned long)pool;
 224                 rwlock_init(&slots->lock);
 225         }
 226
 227         return slots;
 228 }
 229
 230 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 231 {
 232         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 233 }
 234
 235 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 236 {
 237         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 238 }
 239
 240 /* Lock a z3fold page */
 241 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 242 {
 243         spin_lock(&zhdr->page_lock);
 244 }
 245
 246 /* Try to lock a z3fold page */
 247 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 248 {
 249         return spin_trylock(&zhdr->page_lock);
 250 }
 251
 252 /* Unlock a z3fold page */
 253 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 254 {
 255         spin_unlock(&zhdr->page_lock);
 256 }
 257
 258 /* return locked z3fold page if it's not headless */
 259 static inline struct z3fold_header *get_z3fold_header(unsigned long handle)
 260 {
 261         struct z3fold_buddy_slots *slots;
 262         struct z3fold_header *zhdr;
 263         int locked = 0;
 264
 265         if (!(handle & (1 << PAGE_HEADLESS))) {
 266                 slots = handle_to_slots(handle);
 267                 do {
 268                         unsigned long addr;
 269
 270                         read_lock(&slots->lock);
 271                         addr = *(unsigned long *)handle;
 272                         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 273                         locked = z3fold_page_trylock(zhdr);
 274                         read_unlock(&slots->lock);
 275                         if (locked)
 276                                 break;
 277                         cpu_relax();
 278                 } while (true);
 279         } else {
 280                 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
 281         }
 282
 283         return zhdr;
 284 }
 285
 286 static inline void put_z3fold_header(struct z3fold_header *zhdr)
 287 {
 288         struct page *page = virt_to_page(zhdr);
 289
 290         if (!test_bit(PAGE_HEADLESS, &page->private))
 291                 z3fold_page_unlock(zhdr);
 292 }
 293
 294 static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr)
 295 {
 296         struct z3fold_buddy_slots *slots;
 297         int i;
 298         bool is_free;
 299
 300         if (handle & (1 << PAGE_HEADLESS))
 301                 return;
 302
 303         if (WARN_ON(*(unsigned long *)handle == 0))
 304                 return;
 305
 306         slots = handle_to_slots(handle);
 307         write_lock(&slots->lock);
 308         *(unsigned long *)handle = 0;
 309
 310         if (test_bit(HANDLES_NOFREE, &slots->pool)) {
 311                 write_unlock(&slots->lock);
 312                 return; /* simple case, nothing else to do */
 313         }
 314
 315         if (zhdr->slots != slots)
 316                 zhdr->foreign_handles--;
 317
 318         is_free = true;
 319         for (i = 0; i <= BUDDY_MASK; i++) {
 320                 if (slots->slot[i]) {
 321                         is_free = false;
 322                         break;
 323                 }
 324         }
 325         write_unlock(&slots->lock);
 326
 327         if (is_free) {
 328                 struct z3fold_pool *pool = slots_to_pool(slots);
 329
 330                 if (zhdr->slots == slots)
 331                         zhdr->slots = NULL;
 332                 kmem_cache_free(pool->c_handle, slots);
 333         }
 334 }
 335
 336 static int z3fold_init_fs_context(struct fs_context *fc)
 337 {
 338         return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
 339 }
 340
 341 static struct file_system_type z3fold_fs = {
 342         .name           = "z3fold",
 343         .init_fs_context = z3fold_init_fs_context,
 344         .kill_sb        = kill_anon_super,
 345 };
 346
 347 static struct vfsmount *z3fold_mnt;
 348 static int z3fold_mount(void)
 349 {
 350         int ret = 0;
 351
 352         z3fold_mnt = kern_mount(&z3fold_fs);
 353         if (IS_ERR(z3fold_mnt))
 354                 ret = PTR_ERR(z3fold_mnt);
 355
 356         return ret;
 357 }
 358
 359 static void z3fold_unmount(void)
 360 {
 361         kern_unmount(z3fold_mnt);
 362 }
 363
 364 static const struct address_space_operations z3fold_aops;
 365 static int z3fold_register_migration(struct z3fold_pool *pool)
 366 {
 367         pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
 368         if (IS_ERR(pool->inode)) {
 369                 pool->inode = NULL;
 370                 return 1;
 371         }
 372
 373         pool->inode->i_mapping->private_data = pool;
 374         pool->inode->i_mapping->a_ops = &z3fold_aops;
 375         return 0;
 376 }
 377
 378 static void z3fold_unregister_migration(struct z3fold_pool *pool)
 379 {
 380         if (pool->inode)
 381                 iput(pool->inode);
 382 }
 383
 384 /* Initializes the z3fold header of a newly allocated z3fold page */
 385 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 386                                         struct z3fold_pool *pool, gfp_t gfp)
 387 {
 388         struct z3fold_header *zhdr = page_address(page);
 389         struct z3fold_buddy_slots *slots;
 390
 391         INIT_LIST_HEAD(&page->lru);
 392         clear_bit(PAGE_HEADLESS, &page->private);
 393         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 394         clear_bit(NEEDS_COMPACTING, &page->private);
 395         clear_bit(PAGE_STALE, &page->private);
 396         clear_bit(PAGE_CLAIMED, &page->private);
 397         if (headless)
 398                 return zhdr;
 399
 400         slots = alloc_slots(pool, gfp);
 401         if (!slots)
 402                 return NULL;
 403
 404         memset(zhdr, 0, sizeof(*zhdr));
 405         spin_lock_init(&zhdr->page_lock);
 406         kref_init(&zhdr->refcount);
 407         zhdr->cpu = -1;
 408         zhdr->slots = slots;
 409         zhdr->pool = pool;
 410         INIT_LIST_HEAD(&zhdr->buddy);
 411         INIT_WORK(&zhdr->work, compact_page_work);
 412         return zhdr;
 413 }
 414
 415 /* Resets the struct page fields and frees the page */
 416 static void free_z3fold_page(struct page *page, bool headless)
 417 {
 418         if (!headless) {
 419                 lock_page(page);
 420                 __ClearPageMovable(page);
 421                 unlock_page(page);
 422         }
 423         ClearPagePrivate(page);
 424         __free_page(page);
 425 }
 426
 427 /* Helper function to build the index */
 428 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 429 {
 430         return (bud + zhdr->first_num) & BUDDY_MASK;
 431 }
 432
 433 /*
 434  * Encodes the handle of a particular buddy within a z3fold page
 435  * Pool lock should be held as this function accesses first_num
 436  */
 437 static unsigned long __encode_handle(struct z3fold_header *zhdr,
 438                                 struct z3fold_buddy_slots *slots,
 439                                 enum buddy bud)
 440 {
 441         unsigned long h = (unsigned long)zhdr;
 442         int idx = 0;
 443
 444         /*
 445          * For a headless page, its handle is its pointer with the extra
 446          * PAGE_HEADLESS bit set
 447          */
 448         if (bud == HEADLESS)
 449                 return h | (1 << PAGE_HEADLESS);
 450
 451         /* otherwise, return pointer to encoded handle */
 452         idx = __idx(zhdr, bud);
 453         h += idx;
 454         if (bud == LAST)
 455                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 456
 457         write_lock(&slots->lock);
 458         slots->slot[idx] = h;
 459         write_unlock(&slots->lock);
 460         return (unsigned long)&slots->slot[idx];
 461 }
 462
 463 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 464 {
 465         return __encode_handle(zhdr, zhdr->slots, bud);
 466 }
 467
 468 /* only for LAST bud, returns zero otherwise */
 469 static unsigned short handle_to_chunks(unsigned long handle)
 470 {
 471         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 472         unsigned long addr;
 473
 474         read_lock(&slots->lock);
 475         addr = *(unsigned long *)handle;
 476         read_unlock(&slots->lock);
 477         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 478 }
 479
 480 /*
 481  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 482  *  but that doesn't matter. because the masking will result in the
 483  *  correct buddy number.
 484  */
 485 static enum buddy handle_to_buddy(unsigned long handle)
 486 {
 487         struct z3fold_header *zhdr;
 488         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 489         unsigned long addr;
 490
 491         read_lock(&slots->lock);
 492         WARN_ON(handle & (1 << PAGE_HEADLESS));
 493         addr = *(unsigned long *)handle;
 494         read_unlock(&slots->lock);
 495         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 496         return (addr - zhdr->first_num) & BUDDY_MASK;
 497 }
 498
 499 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 500 {
 501         return zhdr->pool;
 502 }
 503
 504 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 505 {
 506         struct page *page = virt_to_page(zhdr);
 507         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 508
 509         WARN_ON(!list_empty(&zhdr->buddy));
 510         set_bit(PAGE_STALE, &page->private);
 511         clear_bit(NEEDS_COMPACTING, &page->private);
 512         spin_lock(&pool->lock);
 513         if (!list_empty(&page->lru))
 514                 list_del_init(&page->lru);
 515         spin_unlock(&pool->lock);
 516
 517         if (locked)
 518                 z3fold_page_unlock(zhdr);
 519
 520         spin_lock(&pool->stale_lock);
 521         list_add(&zhdr->buddy, &pool->stale);
 522         queue_work(pool->release_wq, &pool->work);
 523         spin_unlock(&pool->stale_lock);
 524 }
 525
 526 static void release_z3fold_page(struct kref *ref)
 527 {
 528         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 529                                                 refcount);
 530         __release_z3fold_page(zhdr, false);
 531 }
 532
 533 static void release_z3fold_page_locked(struct kref *ref)
 534 {
 535         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 536                                                 refcount);
 537         WARN_ON(z3fold_page_trylock(zhdr));
 538         __release_z3fold_page(zhdr, true);
 539 }
 540
 541 static void release_z3fold_page_locked_list(struct kref *ref)
 542 {
 543         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 544                                                refcount);
 545         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 546
 547         spin_lock(&pool->lock);
 548         list_del_init(&zhdr->buddy);
 549         spin_unlock(&pool->lock);
 550
 551         WARN_ON(z3fold_page_trylock(zhdr));
 552         __release_z3fold_page(zhdr, true);
 553 }
 554
 555 static void free_pages_work(struct work_struct *w)
 556 {
 557         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 558
 559         spin_lock(&pool->stale_lock);
 560         while (!list_empty(&pool->stale)) {
 561                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 562                                                 struct z3fold_header, buddy);
 563                 struct page *page = virt_to_page(zhdr);
 564
 565                 list_del(&zhdr->buddy);
 566                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 567                         continue;
 568                 spin_unlock(&pool->stale_lock);
 569                 cancel_work_sync(&zhdr->work);
 570                 free_z3fold_page(page, false);
 571                 cond_resched();
 572                 spin_lock(&pool->stale_lock);
 573         }
 574         spin_unlock(&pool->stale_lock);
 575 }
 576
 577 /*
 578  * Returns the number of free chunks in a z3fold page.
 579  * NB: can't be used with HEADLESS pages.
 580  */
 581 static int num_free_chunks(struct z3fold_header *zhdr)
 582 {
 583         int nfree;
 584         /*
 585          * If there is a middle object, pick up the bigger free space
 586          * either before or after it. Otherwise just subtract the number
 587          * of chunks occupied by the first and the last objects.
 588          */
 589         if (zhdr->middle_chunks != 0) {
 590                 int nfree_before = zhdr->first_chunks ?
 591                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 592                 int nfree_after = zhdr->last_chunks ?
 593                         0 : TOTAL_CHUNKS -
 594                                 (zhdr->start_middle + zhdr->middle_chunks);
 595                 nfree = max(nfree_before, nfree_after);
 596         } else
 597                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 598         return nfree;
 599 }
 600
 601 /* Add to the appropriate unbuddied list */
 602 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 603                                 struct z3fold_header *zhdr)
 604 {
 605         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 606                         zhdr->middle_chunks == 0) {
 607                 struct list_head *unbuddied;
 608                 int freechunks = num_free_chunks(zhdr);
 609
 610                 migrate_disable();
 611                 unbuddied = this_cpu_ptr(pool->unbuddied);
 612                 spin_lock(&pool->lock);
 613                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 614                 spin_unlock(&pool->lock);
 615                 zhdr->cpu = smp_processor_id();
 616                 migrate_enable();
 617         }
 618 }
 619
 620 static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks)
 621 {
 622         enum buddy bud = HEADLESS;
 623
 624         if (zhdr->middle_chunks) {
 625                 if (!zhdr->first_chunks &&
 626                     chunks <= zhdr->start_middle - ZHDR_CHUNKS)
 627                         bud = FIRST;
 628                 else if (!zhdr->last_chunks)
 629                         bud = LAST;
 630         } else {
 631                 if (!zhdr->first_chunks)
 632                         bud = FIRST;
 633                 else if (!zhdr->last_chunks)
 634                         bud = LAST;
 635                 else
 636                         bud = MIDDLE;
 637         }
 638
 639         return bud;
 640 }
 641
 642 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 643                                 unsigned short dst_chunk)
 644 {
 645         void *beg = zhdr;
 646         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 647                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 648                        zhdr->middle_chunks << CHUNK_SHIFT);
 649 }
 650
 651 static inline bool buddy_single(struct z3fold_header *zhdr)
 652 {
 653         return !((zhdr->first_chunks && zhdr->middle_chunks) ||
 654                         (zhdr->first_chunks && zhdr->last_chunks) ||
 655                         (zhdr->middle_chunks && zhdr->last_chunks));
 656 }
 657
 658 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
 659 {
 660         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 661         void *p = zhdr;
 662         unsigned long old_handle = 0;
 663         size_t sz = 0;
 664         struct z3fold_header *new_zhdr = NULL;
 665         int first_idx = __idx(zhdr, FIRST);
 666         int middle_idx = __idx(zhdr, MIDDLE);
 667         int last_idx = __idx(zhdr, LAST);
 668         unsigned short *moved_chunks = NULL;
 669
 670         /*
 671          * No need to protect slots here -- all the slots are "local" and
 672          * the page lock is already taken
 673          */
 674         if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
 675                 p += ZHDR_SIZE_ALIGNED;
 676                 sz = zhdr->first_chunks << CHUNK_SHIFT;
 677                 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
 678                 moved_chunks = &zhdr->first_chunks;
 679         } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
 680                 p += zhdr->start_middle << CHUNK_SHIFT;
 681                 sz = zhdr->middle_chunks << CHUNK_SHIFT;
 682                 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
 683                 moved_chunks = &zhdr->middle_chunks;
 684         } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
 685                 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 686                 sz = zhdr->last_chunks << CHUNK_SHIFT;
 687                 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
 688                 moved_chunks = &zhdr->last_chunks;
 689         }
 690
 691         if (sz > 0) {
 692                 enum buddy new_bud = HEADLESS;
 693                 short chunks = size_to_chunks(sz);
 694                 void *q;
 695
 696                 new_zhdr = __z3fold_alloc(pool, sz, false);
 697                 if (!new_zhdr)
 698                         return NULL;
 699
 700                 if (WARN_ON(new_zhdr == zhdr))
 701                         goto out_fail;
 702
 703                 new_bud = get_free_buddy(new_zhdr, chunks);
 704                 q = new_zhdr;
 705                 switch (new_bud) {
 706                 case FIRST:
 707                         new_zhdr->first_chunks = chunks;
 708                         q += ZHDR_SIZE_ALIGNED;
 709                         break;
 710                 case MIDDLE:
 711                         new_zhdr->middle_chunks = chunks;
 712                         new_zhdr->start_middle =
 713                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 714                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 715                         break;
 716                 case LAST:
 717                         new_zhdr->last_chunks = chunks;
 718                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 719                         break;
 720                 default:
 721                         goto out_fail;
 722                 }
 723                 new_zhdr->foreign_handles++;
 724                 memcpy(q, p, sz);
 725                 write_lock(&zhdr->slots->lock);
 726                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 727                         __idx(new_zhdr, new_bud);
 728                 if (new_bud == LAST)
 729                         *(unsigned long *)old_handle |=
 730                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 731                 write_unlock(&zhdr->slots->lock);
 732                 add_to_unbuddied(pool, new_zhdr);
 733                 z3fold_page_unlock(new_zhdr);
 734
 735                 *moved_chunks = 0;
 736         }
 737
 738         return new_zhdr;
 739
 740 out_fail:
 741         if (new_zhdr) {
 742                 if (kref_put(&new_zhdr->refcount, release_z3fold_page_locked))
 743                         atomic64_dec(&pool->pages_nr);
 744                 else {
 745                         add_to_unbuddied(pool, new_zhdr);
 746                         z3fold_page_unlock(new_zhdr);
 747                 }
 748         }
 749         return NULL;
 750
 751 }
 752
 753 #define BIG_CHUNK_GAP   3
 754 /* Has to be called with lock held */
 755 static int z3fold_compact_page(struct z3fold_header *zhdr)
 756 {
 757         struct page *page = virt_to_page(zhdr);
 758
 759         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 760                 return 0; /* can't move middle chunk, it's used */
 761
 762         if (unlikely(PageIsolated(page)))
 763                 return 0;
 764
 765         if (zhdr->middle_chunks == 0)
 766                 return 0; /* nothing to compact */
 767
 768         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 769                 /* move to the beginning */
 770                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 771                 zhdr->first_chunks = zhdr->middle_chunks;
 772                 zhdr->middle_chunks = 0;
 773                 zhdr->start_middle = 0;
 774                 zhdr->first_num++;
 775                 return 1;
 776         }
 777
 778         /*
 779          * moving data is expensive, so let's only do that if
 780          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 781          */
 782         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 783             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 784                         BIG_CHUNK_GAP) {
 785                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 786                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 787                 return 1;
 788         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 789                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 790                                         + zhdr->middle_chunks) >=
 791                         BIG_CHUNK_GAP) {
 792                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 793                         zhdr->middle_chunks;
 794                 mchunk_memmove(zhdr, new_start);
 795                 zhdr->start_middle = new_start;
 796                 return 1;
 797         }
 798
 799         return 0;
 800 }
 801
 802 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 803 {
 804         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 805         struct page *page;
 806
 807         page = virt_to_page(zhdr);
 808         if (locked)
 809                 WARN_ON(z3fold_page_trylock(zhdr));
 810         else
 811                 z3fold_page_lock(zhdr);
 812         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 813                 z3fold_page_unlock(zhdr);
 814                 return;
 815         }
 816         spin_lock(&pool->lock);
 817         list_del_init(&zhdr->buddy);
 818         spin_unlock(&pool->lock);
 819
 820         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 821                 atomic64_dec(&pool->pages_nr);
 822                 return;
 823         }
 824
 825         if (test_bit(PAGE_STALE, &page->private) ||
 826             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 827                 z3fold_page_unlock(zhdr);
 828                 return;
 829         }
 830
 831         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 832             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 833                 if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
 834                         atomic64_dec(&pool->pages_nr);
 835                 else {
 836                         clear_bit(PAGE_CLAIMED, &page->private);
 837                         z3fold_page_unlock(zhdr);
 838                 }
 839                 return;
 840         }
 841
 842         z3fold_compact_page(zhdr);
 843         add_to_unbuddied(pool, zhdr);
 844         clear_bit(PAGE_CLAIMED, &page->private);
 845         z3fold_page_unlock(zhdr);
 846 }
 847
 848 static void compact_page_work(struct work_struct *w)
 849 {
 850         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 851                                                 work);
 852
 853         do_compact_page(zhdr, false);
 854 }
 855
 856 /* returns _locked_ z3fold page header or NULL */
 857 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 858                                                 size_t size, bool can_sleep)
 859 {
 860         struct z3fold_header *zhdr = NULL;
 861         struct page *page;
 862         struct list_head *unbuddied;
 863         int chunks = size_to_chunks(size), i;
 864
 865 lookup:
 866         migrate_disable();
 867         /* First, try to find an unbuddied z3fold page. */
 868         unbuddied = this_cpu_ptr(pool->unbuddied);
 869         for_each_unbuddied_list(i, chunks) {
 870                 struct list_head *l = &unbuddied[i];
 871
 872                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 873                                         struct z3fold_header, buddy);
 874
 875                 if (!zhdr)
 876                         continue;
 877
 878                 /* Re-check under lock. */
 879                 spin_lock(&pool->lock);
 880                 l = &unbuddied[i];
 881                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 882                                                 struct z3fold_header, buddy)) ||
 883                     !z3fold_page_trylock(zhdr)) {
 884                         spin_unlock(&pool->lock);
 885                         zhdr = NULL;
 886                         migrate_enable();
 887                         if (can_sleep)
 888                                 cond_resched();
 889                         goto lookup;
 890                 }
 891                 list_del_init(&zhdr->buddy);
 892                 zhdr->cpu = -1;
 893                 spin_unlock(&pool->lock);
 894
 895                 page = virt_to_page(zhdr);
 896                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 897                     test_bit(PAGE_CLAIMED, &page->private)) {
 898                         z3fold_page_unlock(zhdr);
 899                         zhdr = NULL;
 900                         migrate_enable();
 901                         if (can_sleep)
 902                                 cond_resched();
 903                         goto lookup;
 904                 }
 905
 906                 /*
 907                  * this page could not be removed from its unbuddied
 908                  * list while pool lock was held, and then we've taken
 909                  * page lock so kref_put could not be called before
 910                  * we got here, so it's safe to just call kref_get()
 911                  */
 912                 kref_get(&zhdr->refcount);
 913                 break;
 914         }
 915         migrate_enable();
 916
 917         if (!zhdr) {
 918                 int cpu;
 919
 920                 /* look for _exact_ match on other cpus' lists */
 921                 for_each_online_cpu(cpu) {
 922                         struct list_head *l;
 923
 924                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 925                         spin_lock(&pool->lock);
 926                         l = &unbuddied[chunks];
 927
 928                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 929                                                 struct z3fold_header, buddy);
 930
 931                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 932                                 spin_unlock(&pool->lock);
 933                                 zhdr = NULL;
 934                                 continue;
 935                         }
 936                         list_del_init(&zhdr->buddy);
 937                         zhdr->cpu = -1;
 938                         spin_unlock(&pool->lock);
 939
 940                         page = virt_to_page(zhdr);
 941                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 942                             test_bit(PAGE_CLAIMED, &page->private)) {
 943                                 z3fold_page_unlock(zhdr);
 944                                 zhdr = NULL;
 945                                 if (can_sleep)
 946                                         cond_resched();
 947                                 continue;
 948                         }
 949                         kref_get(&zhdr->refcount);
 950                         break;
 951                 }
 952         }
 953
 954         if (zhdr && !zhdr->slots)
 955                 zhdr->slots = alloc_slots(pool,
 956                                         can_sleep ? GFP_NOIO : GFP_ATOMIC);
 957         return zhdr;
 958 }
 959
 960 /*
 961  * API Functions
 962  */
 963
 964 /**
 965  * z3fold_create_pool() - create a new z3fold pool
 966  * @name:       pool name
 967  * @gfp:        gfp flags when allocating the z3fold pool structure
 968  * @ops:        user-defined operations for the z3fold pool
 969  *
 970  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 971  * failed.
 972  */
 973 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 974                 const struct z3fold_ops *ops)
 975 {
 976         struct z3fold_pool *pool = NULL;
 977         int i, cpu;
 978
 979         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 980         if (!pool)
 981                 goto out;
 982         pool->c_handle = kmem_cache_create("z3fold_handle",
 983                                 sizeof(struct z3fold_buddy_slots),
 984                                 SLOTS_ALIGN, 0, NULL);
 985         if (!pool->c_handle)
 986                 goto out_c;
 987         spin_lock_init(&pool->lock);
 988         spin_lock_init(&pool->stale_lock);
 989         pool->unbuddied = __alloc_percpu(sizeof(struct list_head) * NCHUNKS,
 990                                          __alignof__(struct list_head));
 991         if (!pool->unbuddied)
 992                 goto out_pool;
 993         for_each_possible_cpu(cpu) {
 994                 struct list_head *unbuddied =
 995                                 per_cpu_ptr(pool->unbuddied, cpu);
 996                 for_each_unbuddied_list(i, 0)
 997                         INIT_LIST_HEAD(&unbuddied[i]);
 998         }
 999         INIT_LIST_HEAD(&pool->lru);
1000         INIT_LIST_HEAD(&pool->stale);
1001         atomic64_set(&pool->pages_nr, 0);
1002         pool->name = name;
1003         pool->compact_wq = create_singlethread_workqueue(pool->name);
1004         if (!pool->compact_wq)
1005                 goto out_unbuddied;
1006         pool->release_wq = create_singlethread_workqueue(pool->name);
1007         if (!pool->release_wq)
1008                 goto out_wq;
1009         if (z3fold_register_migration(pool))
1010                 goto out_rwq;
1011         INIT_WORK(&pool->work, free_pages_work);
1012         pool->ops = ops;
1013         return pool;
1014
1015 out_rwq:
1016         destroy_workqueue(pool->release_wq);
1017 out_wq:
1018         destroy_workqueue(pool->compact_wq);
1019 out_unbuddied:
1020         free_percpu(pool->unbuddied);
1021 out_pool:
1022         kmem_cache_destroy(pool->c_handle);
1023 out_c:
1024         kfree(pool);
1025 out:
1026         return NULL;
1027 }
1028
1029 /**
1030  * z3fold_destroy_pool() - destroys an existing z3fold pool
1031  * @pool:       the z3fold pool to be destroyed
1032  *
1033  * The pool should be emptied before this function is called.
1034  */
1035 static void z3fold_destroy_pool(struct z3fold_pool *pool)
1036 {
1037         kmem_cache_destroy(pool->c_handle);
1038
1039         /*
1040          * We need to destroy pool->compact_wq before pool->release_wq,
1041          * as any pending work on pool->compact_wq will call
1042          * queue_work(pool->release_wq, &pool->work).
1043          *
1044          * There are still outstanding pages until both workqueues are drained,
1045          * so we cannot unregister migration until then.
1046          */
1047
1048         destroy_workqueue(pool->compact_wq);
1049         destroy_workqueue(pool->release_wq);
1050         z3fold_unregister_migration(pool);
1051         free_percpu(pool->unbuddied);
1052         kfree(pool);
1053 }
1054
1055 /**
1056  * z3fold_alloc() - allocates a region of a given size
1057  * @pool:       z3fold pool from which to allocate
1058  * @size:       size in bytes of the desired allocation
1059  * @gfp:        gfp flags used if the pool needs to grow
1060  * @handle:     handle of the new allocation
1061  *
1062  * This function will attempt to find a free region in the pool large enough to
1063  * satisfy the allocation request.  A search of the unbuddied lists is
1064  * performed first. If no suitable free region is found, then a new page is
1065  * allocated and added to the pool to satisfy the request.
1066  *
1067  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
1068  * as z3fold pool pages.
1069  *
1070  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
1071  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
1072  * a new page.
1073  */
1074 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1075                         unsigned long *handle)
1076 {
1077         int chunks = size_to_chunks(size);
1078         struct z3fold_header *zhdr = NULL;
1079         struct page *page = NULL;
1080         enum buddy bud;
1081         bool can_sleep = gfpflags_allow_blocking(gfp);
1082
1083         if (!size)
1084                 return -EINVAL;
1085
1086         if (size > PAGE_SIZE)
1087                 return -ENOSPC;
1088
1089         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1090                 bud = HEADLESS;
1091         else {
1092 retry:
1093                 zhdr = __z3fold_alloc(pool, size, can_sleep);
1094                 if (zhdr) {
1095                         bud = get_free_buddy(zhdr, chunks);
1096                         if (bud == HEADLESS) {
1097                                 if (kref_put(&zhdr->refcount,
1098                                              release_z3fold_page_locked))
1099                                         atomic64_dec(&pool->pages_nr);
1100                                 else
1101                                         z3fold_page_unlock(zhdr);
1102                                 pr_err("No free chunks in unbuddied\n");
1103                                 WARN_ON(1);
1104                                 goto retry;
1105                         }
1106                         page = virt_to_page(zhdr);
1107                         goto found;
1108                 }
1109                 bud = FIRST;
1110         }
1111
1112         page = NULL;
1113         if (can_sleep) {
1114                 spin_lock(&pool->stale_lock);
1115                 zhdr = list_first_entry_or_null(&pool->stale,
1116                                                 struct z3fold_header, buddy);
1117                 /*
1118                  * Before allocating a page, let's see if we can take one from
1119                  * the stale pages list. cancel_work_sync() can sleep so we
1120                  * limit this case to the contexts where we can sleep
1121                  */
1122                 if (zhdr) {
1123                         list_del(&zhdr->buddy);
1124                         spin_unlock(&pool->stale_lock);
1125                         cancel_work_sync(&zhdr->work);
1126                         page = virt_to_page(zhdr);
1127                 } else {
1128                         spin_unlock(&pool->stale_lock);
1129                 }
1130         }
1131         if (!page)
1132                 page = alloc_page(gfp);
1133
1134         if (!page)
1135                 return -ENOMEM;
1136
1137         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1138         if (!zhdr) {
1139                 __free_page(page);
1140                 return -ENOMEM;
1141         }
1142         atomic64_inc(&pool->pages_nr);
1143
1144         if (bud == HEADLESS) {
1145                 set_bit(PAGE_HEADLESS, &page->private);
1146                 goto headless;
1147         }
1148         if (can_sleep) {
1149                 lock_page(page);
1150                 __SetPageMovable(page, pool->inode->i_mapping);
1151                 unlock_page(page);
1152         } else {
1153                 if (trylock_page(page)) {
1154                         __SetPageMovable(page, pool->inode->i_mapping);
1155                         unlock_page(page);
1156                 }
1157         }
1158         z3fold_page_lock(zhdr);
1159
1160 found:
1161         if (bud == FIRST)
1162                 zhdr->first_chunks = chunks;
1163         else if (bud == LAST)
1164                 zhdr->last_chunks = chunks;
1165         else {
1166                 zhdr->middle_chunks = chunks;
1167                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1168         }
1169         add_to_unbuddied(pool, zhdr);
1170
1171 headless:
1172         spin_lock(&pool->lock);
1173         /* Add/move z3fold page to beginning of LRU */
1174         if (!list_empty(&page->lru))
1175                 list_del(&page->lru);
1176
1177         list_add(&page->lru, &pool->lru);
1178
1179         *handle = encode_handle(zhdr, bud);
1180         spin_unlock(&pool->lock);
1181         if (bud != HEADLESS)
1182                 z3fold_page_unlock(zhdr);
1183
1184         return 0;
1185 }
1186
1187 /**
1188  * z3fold_free() - frees the allocation associated with the given handle
1189  * @pool:       pool in which the allocation resided
1190  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1191  *
1192  * In the case that the z3fold page in which the allocation resides is under
1193  * reclaim, as indicated by the PG_reclaim flag being set, this function
1194  * only sets the first|last_chunks to 0.  The page is actually freed
1195  * once both buddies are evicted (see z3fold_reclaim_page() below).
1196  */
1197 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1198 {
1199         struct z3fold_header *zhdr;
1200         struct page *page;
1201         enum buddy bud;
1202         bool page_claimed;
1203
1204         zhdr = get_z3fold_header(handle);
1205         page = virt_to_page(zhdr);
1206         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1207
1208         if (test_bit(PAGE_HEADLESS, &page->private)) {
1209                 /* if a headless page is under reclaim, just leave.
1210                  * NB: we use test_and_set_bit for a reason: if the bit
1211                  * has not been set before, we release this page
1212                  * immediately so we don't care about its value any more.
1213                  */
1214                 if (!page_claimed) {
1215                         spin_lock(&pool->lock);
1216                         list_del(&page->lru);
1217                         spin_unlock(&pool->lock);
1218                         put_z3fold_header(zhdr);
1219                         free_z3fold_page(page, true);
1220                         atomic64_dec(&pool->pages_nr);
1221                 }
1222                 return;
1223         }
1224
1225         /* Non-headless case */
1226         bud = handle_to_buddy(handle);
1227
1228         switch (bud) {
1229         case FIRST:
1230                 zhdr->first_chunks = 0;
1231                 break;
1232         case MIDDLE:
1233                 zhdr->middle_chunks = 0;
1234                 break;
1235         case LAST:
1236                 zhdr->last_chunks = 0;
1237                 break;
1238         default:
1239                 pr_err("%s: unknown bud %d\n", __func__, bud);
1240                 WARN_ON(1);
1241                 put_z3fold_header(zhdr);
1242                 return;
1243         }
1244
1245         if (!page_claimed)
1246                 free_handle(handle, zhdr);
1247         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1248                 atomic64_dec(&pool->pages_nr);
1249                 return;
1250         }
1251         if (page_claimed) {
1252                 /* the page has not been claimed by us */
1253                 z3fold_page_unlock(zhdr);
1254                 return;
1255         }
1256         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1257                 put_z3fold_header(zhdr);
1258                 clear_bit(PAGE_CLAIMED, &page->private);
1259                 return;
1260         }
1261         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1262                 spin_lock(&pool->lock);
1263                 list_del_init(&zhdr->buddy);
1264                 spin_unlock(&pool->lock);
1265                 zhdr->cpu = -1;
1266                 kref_get(&zhdr->refcount);
1267                 clear_bit(PAGE_CLAIMED, &page->private);
1268                 do_compact_page(zhdr, true);
1269                 return;
1270         }
1271         kref_get(&zhdr->refcount);
1272         clear_bit(PAGE_CLAIMED, &page->private);
1273         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1274         put_z3fold_header(zhdr);
1275 }
1276
1277 /**
1278  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1279  * @pool:       pool from which a page will attempt to be evicted
1280  * @retries:    number of pages on the LRU list for which eviction will
1281  *              be attempted before failing
1282  *
1283  * z3fold reclaim is different from normal system reclaim in that it is done
1284  * from the bottom, up. This is because only the bottom layer, z3fold, has
1285  * information on how the allocations are organized within each z3fold page.
1286  * This has the potential to create interesting locking situations between
1287  * z3fold and the user, however.
1288  *
1289  * To avoid these, this is how z3fold_reclaim_page() should be called:
1290  *
1291  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1292  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1293  * call the user-defined eviction handler with the pool and handle as
1294  * arguments.
1295  *
1296  * If the handle can not be evicted, the eviction handler should return
1297  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1298  * appropriate list and try the next z3fold page on the LRU up to
1299  * a user defined number of retries.
1300  *
1301  * If the handle is successfully evicted, the eviction handler should
1302  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1303  * contains logic to delay freeing the page if the page is under reclaim,
1304  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1305  *
1306  * If all buddies in the z3fold page are successfully evicted, then the
1307  * z3fold page can be freed.
1308  *
1309  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1310  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1311  * the retry limit was hit.
1312  */
1313 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1314 {
1315         int i, ret = -1;
1316         struct z3fold_header *zhdr = NULL;
1317         struct page *page = NULL;
1318         struct list_head *pos;
1319         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1320         struct z3fold_buddy_slots slots __attribute__((aligned(SLOTS_ALIGN)));
1321
1322         rwlock_init(&slots.lock);
1323         slots.pool = (unsigned long)pool | (1 << HANDLES_NOFREE);
1324
1325         spin_lock(&pool->lock);
1326         if (!pool->ops || !pool->ops->evict || retries == 0) {
1327                 spin_unlock(&pool->lock);
1328                 return -EINVAL;
1329         }
1330         for (i = 0; i < retries; i++) {
1331                 if (list_empty(&pool->lru)) {
1332                         spin_unlock(&pool->lock);
1333                         return -EINVAL;
1334                 }
1335                 list_for_each_prev(pos, &pool->lru) {
1336                         page = list_entry(pos, struct page, lru);
1337
1338                         zhdr = page_address(page);
1339                         if (test_bit(PAGE_HEADLESS, &page->private)) {
1340                                 /*
1341                                  * For non-headless pages, we wait to do this
1342                                  * until we have the page lock to avoid racing
1343                                  * with __z3fold_alloc(). Headless pages don't
1344                                  * have a lock (and __z3fold_alloc() will never
1345                                  * see them), but we still need to test and set
1346                                  * PAGE_CLAIMED to avoid racing with
1347                                  * z3fold_free(), so just do it now before
1348                                  * leaving the loop.
1349                                  */
1350                                 if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1351                                         continue;
1352
1353                                 break;
1354                         }
1355
1356                         if (kref_get_unless_zero(&zhdr->refcount) == 0) {
1357                                 zhdr = NULL;
1358                                 break;
1359                         }
1360                         if (!z3fold_page_trylock(zhdr)) {
1361                                 if (kref_put(&zhdr->refcount,
1362                                                 release_z3fold_page))
1363                                         atomic64_dec(&pool->pages_nr);
1364                                 zhdr = NULL;
1365                                 continue; /* can't evict at this point */
1366                         }
1367
1368                         /* test_and_set_bit is of course atomic, but we still
1369                          * need to do it under page lock, otherwise checking
1370                          * that bit in __z3fold_alloc wouldn't make sense
1371                          */
1372                         if (zhdr->foreign_handles ||
1373                             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1374                                 if (kref_put(&zhdr->refcount,
1375                                                 release_z3fold_page_locked))
1376                                         atomic64_dec(&pool->pages_nr);
1377                                 else
1378                                         z3fold_page_unlock(zhdr);
1379                                 zhdr = NULL;
1380                                 continue; /* can't evict such page */
1381                         }
1382                         list_del_init(&zhdr->buddy);
1383                         zhdr->cpu = -1;
1384                         break;
1385                 }
1386
1387                 if (!zhdr)
1388                         break;
1389
1390                 list_del_init(&page->lru);
1391                 spin_unlock(&pool->lock);
1392
1393                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1394                         /*
1395                          * We need encode the handles before unlocking, and
1396                          * use our local slots structure because z3fold_free
1397                          * can zero out zhdr->slots and we can't do much
1398                          * about that
1399                          */
1400                         first_handle = 0;
1401                         last_handle = 0;
1402                         middle_handle = 0;
1403                         memset(slots.slot, 0, sizeof(slots.slot));
1404                         if (zhdr->first_chunks)
1405                                 first_handle = __encode_handle(zhdr, &slots,
1406                                                                 FIRST);
1407                         if (zhdr->middle_chunks)
1408                                 middle_handle = __encode_handle(zhdr, &slots,
1409                                                                 MIDDLE);
1410                         if (zhdr->last_chunks)
1411                                 last_handle = __encode_handle(zhdr, &slots,
1412                                                                 LAST);
1413                         /*
1414                          * it's safe to unlock here because we hold a
1415                          * reference to this page
1416                          */
1417                         z3fold_page_unlock(zhdr);
1418                 } else {
1419                         first_handle = encode_handle(zhdr, HEADLESS);
1420                         last_handle = middle_handle = 0;
1421                 }
1422                 /* Issue the eviction callback(s) */
1423                 if (middle_handle) {
1424                         ret = pool->ops->evict(pool, middle_handle);
1425                         if (ret)
1426                                 goto next;
1427                 }
1428                 if (first_handle) {
1429                         ret = pool->ops->evict(pool, first_handle);
1430                         if (ret)
1431                                 goto next;
1432                 }
1433                 if (last_handle) {
1434                         ret = pool->ops->evict(pool, last_handle);
1435                         if (ret)
1436                                 goto next;
1437                 }
1438 next:
1439                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1440                         if (ret == 0) {
1441                                 free_z3fold_page(page, true);
1442                                 atomic64_dec(&pool->pages_nr);
1443                                 return 0;
1444                         }
1445                         spin_lock(&pool->lock);
1446                         list_add(&page->lru, &pool->lru);
1447                         spin_unlock(&pool->lock);
1448                         clear_bit(PAGE_CLAIMED, &page->private);
1449                 } else {
1450                         struct z3fold_buddy_slots *slots = zhdr->slots;
1451                         z3fold_page_lock(zhdr);
1452                         if (kref_put(&zhdr->refcount,
1453                                         release_z3fold_page_locked)) {
1454                                 kmem_cache_free(pool->c_handle, slots);
1455                                 atomic64_dec(&pool->pages_nr);
1456                                 return 0;
1457                         }
1458                         /*
1459                          * if we are here, the page is still not completely
1460                          * free. Take the global pool lock then to be able
1461                          * to add it back to the lru list
1462                          */
1463                         spin_lock(&pool->lock);
1464                         list_add(&page->lru, &pool->lru);
1465                         spin_unlock(&pool->lock);
1466                         z3fold_page_unlock(zhdr);
1467                         clear_bit(PAGE_CLAIMED, &page->private);
1468                 }
1469
1470                 /* We started off locked to we need to lock the pool back */
1471                 spin_lock(&pool->lock);
1472         }
1473         spin_unlock(&pool->lock);
1474         return -EAGAIN;
1475 }
1476
1477 /**
1478  * z3fold_map() - maps the allocation associated with the given handle
1479  * @pool:       pool in which the allocation resides
1480  * @handle:     handle associated with the allocation to be mapped
1481  *
1482  * Extracts the buddy number from handle and constructs the pointer to the
1483  * correct starting chunk within the page.
1484  *
1485  * Returns: a pointer to the mapped allocation
1486  */
1487 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1488 {
1489         struct z3fold_header *zhdr;
1490         struct page *page;
1491         void *addr;
1492         enum buddy buddy;
1493
1494         zhdr = get_z3fold_header(handle);
1495         addr = zhdr;
1496         page = virt_to_page(zhdr);
1497
1498         if (test_bit(PAGE_HEADLESS, &page->private))
1499                 goto out;
1500
1501         buddy = handle_to_buddy(handle);
1502         switch (buddy) {
1503         case FIRST:
1504                 addr += ZHDR_SIZE_ALIGNED;
1505                 break;
1506         case MIDDLE:
1507                 addr += zhdr->start_middle << CHUNK_SHIFT;
1508                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1509                 break;
1510         case LAST:
1511                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1512                 break;
1513         default:
1514                 pr_err("unknown buddy id %d\n", buddy);
1515                 WARN_ON(1);
1516                 addr = NULL;
1517                 break;
1518         }
1519
1520         if (addr)
1521                 zhdr->mapped_count++;
1522 out:
1523         put_z3fold_header(zhdr);
1524         return addr;
1525 }
1526
1527 /**
1528  * z3fold_unmap() - unmaps the allocation associated with the given handle
1529  * @pool:       pool in which the allocation resides
1530  * @handle:     handle associated with the allocation to be unmapped
1531  */
1532 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1533 {
1534         struct z3fold_header *zhdr;
1535         struct page *page;
1536         enum buddy buddy;
1537
1538         zhdr = get_z3fold_header(handle);
1539         page = virt_to_page(zhdr);
1540
1541         if (test_bit(PAGE_HEADLESS, &page->private))
1542                 return;
1543
1544         buddy = handle_to_buddy(handle);
1545         if (buddy == MIDDLE)
1546                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1547         zhdr->mapped_count--;
1548         put_z3fold_header(zhdr);
1549 }
1550
1551 /**
1552  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1553  * @pool:       pool whose size is being queried
1554  *
1555  * Returns: size in pages of the given pool.
1556  */
1557 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1558 {
1559         return atomic64_read(&pool->pages_nr);
1560 }
1561
1562 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1563 {
1564         struct z3fold_header *zhdr;
1565         struct z3fold_pool *pool;
1566
1567         VM_BUG_ON_PAGE(!PageMovable(page), page);
1568         VM_BUG_ON_PAGE(PageIsolated(page), page);
1569
1570         if (test_bit(PAGE_HEADLESS, &page->private))
1571                 return false;
1572
1573         zhdr = page_address(page);
1574         z3fold_page_lock(zhdr);
1575         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1576             test_bit(PAGE_STALE, &page->private))
1577                 goto out;
1578
1579         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1580                 goto out;
1581
1582         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1583                 goto out;
1584         pool = zhdr_to_pool(zhdr);
1585         spin_lock(&pool->lock);
1586         if (!list_empty(&zhdr->buddy))
1587                 list_del_init(&zhdr->buddy);
1588         if (!list_empty(&page->lru))
1589                 list_del_init(&page->lru);
1590         spin_unlock(&pool->lock);
1591
1592         kref_get(&zhdr->refcount);
1593         z3fold_page_unlock(zhdr);
1594         return true;
1595
1596 out:
1597         z3fold_page_unlock(zhdr);
1598         return false;
1599 }
1600
1601 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1602                                struct page *page, enum migrate_mode mode)
1603 {
1604         struct z3fold_header *zhdr, *new_zhdr;
1605         struct z3fold_pool *pool;
1606         struct address_space *new_mapping;
1607
1608         VM_BUG_ON_PAGE(!PageMovable(page), page);
1609         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1610         VM_BUG_ON_PAGE(!test_bit(PAGE_CLAIMED, &page->private), page);
1611         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1612
1613         zhdr = page_address(page);
1614         pool = zhdr_to_pool(zhdr);
1615
1616         if (!z3fold_page_trylock(zhdr))
1617                 return -EAGAIN;
1618         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1619                 z3fold_page_unlock(zhdr);
1620                 clear_bit(PAGE_CLAIMED, &page->private);
1621                 return -EBUSY;
1622         }
1623         if (work_pending(&zhdr->work)) {
1624                 z3fold_page_unlock(zhdr);
1625                 return -EAGAIN;
1626         }
1627         new_zhdr = page_address(newpage);
1628         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1629         newpage->private = page->private;
1630         page->private = 0;
1631         z3fold_page_unlock(zhdr);
1632         spin_lock_init(&new_zhdr->page_lock);
1633         INIT_WORK(&new_zhdr->work, compact_page_work);
1634         /*
1635          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1636          * so we only have to reinitialize it.
1637          */
1638         INIT_LIST_HEAD(&new_zhdr->buddy);
1639         new_mapping = page_mapping(page);
1640         __ClearPageMovable(page);
1641         ClearPagePrivate(page);
1642
1643         get_page(newpage);
1644         z3fold_page_lock(new_zhdr);
1645         if (new_zhdr->first_chunks)
1646                 encode_handle(new_zhdr, FIRST);
1647         if (new_zhdr->last_chunks)
1648                 encode_handle(new_zhdr, LAST);
1649         if (new_zhdr->middle_chunks)
1650                 encode_handle(new_zhdr, MIDDLE);
1651         set_bit(NEEDS_COMPACTING, &newpage->private);
1652         new_zhdr->cpu = smp_processor_id();
1653         spin_lock(&pool->lock);
1654         list_add(&newpage->lru, &pool->lru);
1655         spin_unlock(&pool->lock);
1656         __SetPageMovable(newpage, new_mapping);
1657         z3fold_page_unlock(new_zhdr);
1658
1659         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1660
1661         page_mapcount_reset(page);
1662         clear_bit(PAGE_CLAIMED, &page->private);
1663         put_page(page);
1664         return 0;
1665 }
1666
1667 static void z3fold_page_putback(struct page *page)
1668 {
1669         struct z3fold_header *zhdr;
1670         struct z3fold_pool *pool;
1671
1672         zhdr = page_address(page);
1673         pool = zhdr_to_pool(zhdr);
1674
1675         z3fold_page_lock(zhdr);
1676         if (!list_empty(&zhdr->buddy))
1677                 list_del_init(&zhdr->buddy);
1678         INIT_LIST_HEAD(&page->lru);
1679         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1680                 atomic64_dec(&pool->pages_nr);
1681                 return;
1682         }
1683         spin_lock(&pool->lock);
1684         list_add(&page->lru, &pool->lru);
1685         spin_unlock(&pool->lock);
1686         clear_bit(PAGE_CLAIMED, &page->private);
1687         z3fold_page_unlock(zhdr);
1688 }
1689
1690 static const struct address_space_operations z3fold_aops = {
1691         .isolate_page = z3fold_page_isolate,
1692         .migratepage = z3fold_page_migrate,
1693         .putback_page = z3fold_page_putback,
1694 };
1695
1696 /*****************
1697  * zpool
1698  ****************/
1699
1700 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1701 {
1702         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1703                 return pool->zpool_ops->evict(pool->zpool, handle);
1704         else
1705                 return -ENOENT;
1706 }
1707
1708 static const struct z3fold_ops z3fold_zpool_ops = {
1709         .evict =        z3fold_zpool_evict
1710 };
1711
1712 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1713                                const struct zpool_ops *zpool_ops,
1714                                struct zpool *zpool)
1715 {
1716         struct z3fold_pool *pool;
1717
1718         pool = z3fold_create_pool(name, gfp,
1719                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1720         if (pool) {
1721                 pool->zpool = zpool;
1722                 pool->zpool_ops = zpool_ops;
1723         }
1724         return pool;
1725 }
1726
1727 static void z3fold_zpool_destroy(void *pool)
1728 {
1729         z3fold_destroy_pool(pool);
1730 }
1731
1732 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1733                         unsigned long *handle)
1734 {
1735         return z3fold_alloc(pool, size, gfp, handle);
1736 }
1737 static void z3fold_zpool_free(void *pool, unsigned long handle)
1738 {
1739         z3fold_free(pool, handle);
1740 }
1741
1742 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1743                         unsigned int *reclaimed)
1744 {
1745         unsigned int total = 0;
1746         int ret = -EINVAL;
1747
1748         while (total < pages) {
1749                 ret = z3fold_reclaim_page(pool, 8);
1750                 if (ret < 0)
1751                         break;
1752                 total++;
1753         }
1754
1755         if (reclaimed)
1756                 *reclaimed = total;
1757
1758         return ret;
1759 }
1760
1761 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1762                         enum zpool_mapmode mm)
1763 {
1764         return z3fold_map(pool, handle);
1765 }
1766 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1767 {
1768         z3fold_unmap(pool, handle);
1769 }
1770
1771 static u64 z3fold_zpool_total_size(void *pool)
1772 {
1773         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1774 }
1775
1776 static struct zpool_driver z3fold_zpool_driver = {
1777         .type =         "z3fold",
1778         .sleep_mapped = true,
1779         .owner =        THIS_MODULE,
1780         .create =       z3fold_zpool_create,
1781         .destroy =      z3fold_zpool_destroy,
1782         .malloc =       z3fold_zpool_malloc,
1783         .free =         z3fold_zpool_free,
1784         .shrink =       z3fold_zpool_shrink,
1785         .map =          z3fold_zpool_map,
1786         .unmap =        z3fold_zpool_unmap,
1787         .total_size =   z3fold_zpool_total_size,
1788 };
1789
1790 MODULE_ALIAS("zpool-z3fold");
1791
1792 static int __init init_z3fold(void)
1793 {
1794         int ret;
1795
1796         /*
1797          * Make sure the z3fold header is not larger than the page size and
1798          * there has remaining spaces for its buddy.
1799          */
1800         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE - CHUNK_SIZE);
1801         ret = z3fold_mount();
1802         if (ret)
1803                 return ret;
1804
1805         zpool_register_driver(&z3fold_zpool_driver);
1806
1807         return 0;
1808 }
1809
1810 static void __exit exit_z3fold(void)
1811 {
1812         z3fold_unmount();
1813         zpool_unregister_driver(&z3fold_zpool_driver);
1814 }
1815
1816 module_init(init_z3fold);
1817 module_exit(exit_z3fold);
1818
1819 MODULE_LICENSE("GPL");
1820 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1821 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");